MATHEMATICS COLLOQUIUM

 WHEN:
On Thursday, September 27, 2001 at 2.30PM

WHERE:
Durham Science Center, Room 255

WHAT:

Todd Munson
Argonne National Laboratories

will give a talk on

Two Support Vector Machine Algorithms

Abstract:
We discuss interior-point and semismooth methods for solving quadratic programming problems with a small number of linear constraints where the quadratic term consists of a low-rank update to a positive semi-definite matrix. Several formulations of the support vector machine, a technique employed by the machine learning community for supervised learning, fit into this category. A related example is the Huber regression problem which can also be posed as a quadratic program with the desired properties.

Support vector machines can be used, for example, when evaluating credit card applications to determine whether or not a particular applicant should be approved. An interesting feature of these problems is the volume of data, which can lead to quadratic programs with between 10 and 100 million variables and, if written explicitly, a dense Q matrix. The implementations of the two algorithms use linear algebra specialized for the support vector machine application. For the targeted massive problems, all data is stored out-of-core and we overlap computation and I/O to reduce overhead. Results will be reported for several linear support vector machine formulations demonstrating that the algorithms are reliable and scalable.

This talk is joint work with Michael Ferris.